How Many AI Personas Do You Need? What a Five-Persona Panel Actually Preserves

At a glance

Research snapshot

Research question

Why this question matters

Persona panels are easy to expand.

A research team can ask the same question as a CMO, an SEO Manager, a Founder, a Product Marketing leader, a Demand Generation leader, and then keep adding roles.

At some point, though, more personas become expensive.

They increase model calls, review volume, annotation work, reporting complexity and repeated or redundant outputs.

That creates a practical design question:

How many personas are enough?

The obvious shortcut is to choose two or three roles that seem different and assume they capture most of the useful variation.

But that assumption needs testing.

Our parent study found that five persona-conditioned responses remained strongly distinguishable after direct persona wording was removed. The same persona-associated differentiation also appeared in grounded search-query formulation and remained broadly high across five task families.

This analysis asks the next practical question:

If we reduce the five-persona panel, how much of the observed answer diversity disappears?

For the broader experiment and its causal limitations, see Persona Prompts Change More Than Tone: Evidence from 750 Grounded AI Responses.

Panel design

The five-persona panel

The panel contained:

• CMO
• VP of Product Marketing
• SEO Manager
• Demand Generation Director
• Founder/CEO

Each of 150 neutral scenarios was originally rendered under all five persona conditions.

After the historical quality screen removed one malformed SEO Manager response, the primary panel analysis used 149 complete scenarios.

The scenarios covered five task types:

1. Action plan
2. Comparison
3. Diagnostic
4. Measurement
5. Risk control

The goal here was not to decide which persona gave the best answer.

Instead, we measured how much of the full five-persona response space remained when one or more persona conditions were removed.

Finding 1

Two personas preserved only about 41% of the five-persona answer space

The strongest two-persona subsets retained approximately 41% of the normalized five-persona representation.

That result was strikingly stable across task types.

SEO Manager appeared in every best-performing two-persona subset.

That aligns with the parent study, where SEO Manager was the most distinctive answer condition across all five task families.

But the result should not be interpreted as:

“Always include SEO.”

SEO Manager's inclusion here reflects representational distinctiveness under this experimental design.

It does not establish that SEO answers are more correct, more useful or more valuable.

Open full-resolution figure

Finding 2

Three personas preserved about 61%—well short of near-full coverage

Adding a third persona improved coverage substantially.

But the best three-persona panels still retained only about 61% of the full five-persona answer-space representation.

No best three-persona task panel reached the predefined 80% preservation threshold.

That matters because three-persona research panels are intuitively appealing.

They feel broad enough to capture different viewpoints without becoming expensive.

In this experiment, however, three personas still left a large amount of the five-persona representational space uncovered.

Finding 3

Four personas reached about 81%, but still did not approximate the full panel

Removing only one persona produced a much smaller loss.

The best four-persona subsets preserved approximately 81% of the full-panel response-space geometry.

Four-persona panels passed the 80% preservation level.

But they remained below the separately predefined 95% near-full-coverage criterion.

That led to the study's panel conclusion:

The five-persona panel was materially additive for observed response-space diversity.

Again, “materially additive” refers to geometric coverage.

It does not establish incremental business utility.

Persona distinctiveness

Why does SEO Manager appear so often?

SEO Manager was the most distinctive answer condition for all five task families in the parent analysis.

Removing SEO reduced remaining-pair diversity more than removing any other persona:

By contrast, removing CMO or Founder/CEO slightly increased the mean diversity among the remaining personas.

That does not mean removing CMO or Founder improves the panel.

The denominator changes when a persona is removed, so the remaining set can become more spread out even while losing coverage.

This is why the study measured both leave-one-out diversity and coverage of the original five-persona space.

The second metric is the more useful one for the panel-size question.

Open full-resolution figure

Persona proximity

The closest pair was CMO and Founder/CEO

Across every task family, the nearest persona pair was:

CMO / Founder/CEO

They were consistently more similar to one another than other persona pairs.

But “closest” is relative.

A distance around 0.82 still reflects substantial separation in the residual answer representation.

And lexical proximity does not establish substantive redundancy.

CMO and Founder/CEO could still contribute different factual assertions, risks or recommendations.

That substantive question remains under human validation in the next stage of the research programme.

Finding 4

The panel-size result was broadly stable across task types

One possible explanation was that smaller panels might be sufficient for some tasks but not others.

The result was more stable than that.

The parent analysis found persona differentiation remained high across action plans, comparisons, diagnostics, measurement and risk control.

Panel retention showed a similarly consistent pattern.

Two-persona panels were around 41%.

Three-persona panels were around 61%.

Four-persona panels were around 81%.

That consistency is important.

It suggests the panel-size result was not simply the consequence of one unusually persona-sensitive task family.

Held-out-cluster checks produced similar results.

The study therefore did not identify a task where a small fixed panel suddenly became a near-complete substitute for the five-persona panel.

Open full-resolution figure

Research practice

What should AI research teams do with this?

The wrong conclusion is:

“Everyone should use five personas.”

The useful conclusion is:

Choose persona-panel size based on the coverage objective you are willing to sacrifice.

If the purpose is fast directional research, two or three personas may be enough.

If the purpose is to preserve most of the representational range observed in a broader persona panel, the trade-off becomes much sharper.

A practical workflow is:

1. Define why each persona exists

   Each persona should represent a distinct decision context—not a cosmetic job title.

2. Test redundancy empirically

   Do not assume two executive roles are interchangeable because their titles seem similar. Do not assume a specialist role is useful simply because its vocabulary is different.

3. Separate diversity from quality

   A more diverse panel may expose more answer formulations while still producing redundant substantive recommendations.

4. Start broad, then prune

   For exploratory studies, a broader pilot panel can reveal which conditions are genuinely distinctive. Later runs can reduce the panel if validated coverage remains acceptable.

5. Use a holdout check

   A panel selected on one topic may fail on another. The parent study therefore evaluated subset selection on four topic clusters and tested it on the held-out fifth.

Open question

The unresolved question: representational coverage versus substantive coverage

This is now the most important open question.

Suppose a three-persona panel preserves only 61% of the observed lexical/structural response geometry.

That sounds like a large loss.

But what if the omitted 39% consists mostly of alternative phrasing, different formatting, role-specific terminology or redundant recommendations?

Then the smaller panel may still preserve most of the useful content.

The reverse is also possible.

A persona that looks lexically close to another may contribute a unique factual assertion, strategic recommendation, risk or decision priority.

Kojable's next validation layer is therefore measuring substantive divergence across factual claims, recommendations, risks and decisions.

That analysis is currently awaiting independent human calibration.

Until it is complete, the correct distinction is:

Five personas materially increased observed representational coverage. Whether they materially increase substantive decision coverage is not yet established.

Interpretation

Implications for AI visibility and Answer Alignment research

Persona panels are particularly relevant when testing how AI systems represent companies.

A buyer may approach the same company through different decision contexts:

• Strategic fit
• Technical implementation
• Revenue impact
• Positioning
• Risk

If an evaluation tracks only one generic version of the question, it may under-sample the range of representations the model can produce.

But blindly multiplying prompts is not an answer either.

A useful persona panel should balance coverage, cost, review burden, redundancy and substantive value.

The practical challenge is not to maximize persona count.

It is to identify the smallest panel that preserves the dimensions of AI representation that matter for the research objective.

This study measures one part of that problem: representational coverage.

Conclusion

Conclusion

Persona-panel design is a coverage problem, not a counting exercise.

In this matched five-persona experiment, two personas captured roughly 41% of the observed response-space representation.

Three captured about 61%.

Four captured about 81%.

None of the smaller fixed panels approximated the full five-persona reference under the study's predefined near-full-coverage criterion.

That does not establish that five personas are universally necessary.

It does show that assuming a small persona panel captures “most perspectives” can be wrong even when the underlying business scenario is held constant.

The practical lesson is:

Do not choose persona count by intuition alone. Define the coverage you care about, measure redundancy, and reduce the panel only when the omitted conditions stop contributing dimensions that matter to the research objective.

For the full research programme—including prompt-echo testing, grounded-query differentiation and task robustness—see Persona Prompts Change More Than Tone: Evidence from 750 Grounded AI Responses.

Research context

Related Kojable research

• Persona Prompts Change More Than Tone: Evidence from 750 Grounded AI Responses

  The capstone report for the matched V2 persona programme.

• How Persona Conditioning Changes AEO Answers

  The earlier matched 750-response AEO analysis.

• Beyond Persona Labels: Do AI Personas Still Differ After Explicit Cues Are Removed?

  Separate cue-ablation research.

• When Do Personas Really Change AI Responses?

  Earlier persona-similarity research.

• Do AI Role Prompts Change the Decision Lens?

  Separate marketing-role framing analysis.

• AI Answer and Source Research

  Browse the Kojable Research index.